Epilepsy EEG Diagnosis Classification Based On Ensemble Learning Algorithm

Epilepsy,a chronic brain disease that seriously threatens human health,is caused by abnormal electrical discharge in a group of brain neuron cells.At present,the global patient has reached 50 million and is growing at a rate of 10%,therefore,early medical diagnosis of epilepsy is very important.Epilepsy detection research based on EEG signal,which can reflect the brain activities caused by epilepsy has gradually become the scientific focus.However,the recognition algorithm of epilepsy EEG signals still had several drawbacks such as insufficient collection features,redundant calculations,and weak generalization ability.To solve the above problems,the classification was completed by an integrated learning algorithm in this thesis through several steps,including pretreatment noise reduction,multi-feature acquisition,and dimensionality reduction.The main research work is as follows:(1)Preprocessing of EEG signals.The purpose of preprocessing is to reduce the noise mixed in the EEG signal.The advantages of wavelet threshold denoising and empirical mode decomposition(EMD)denoising were both considered.By improving the threshold function,EMD-wavelet threshold joint denoising method was performed in two states of eye-open and eyeclosed.In the simulations,three kinds of noise reduction methods were comprehensively compared and evaluated with multiple noise-cancellation indexes.The superiority of the EMD-wavelet threshold joint denoising method was verified.(2)Feature extraction.A comprehensive feature extraction of epilepsy EEG signals was carried out.Firstly,6 and 5 commonly used features were extracted respectively in the time domain and frequency domain.By combining with the frequency characteristics of typical epilepsy waves,6 features were extracted by wavelet decomposition in the time-frequency domain.Secondly,the sample entropy was used as the nonlinear feature to describe the signal confusion.Finally,the rhythm feature of epilepsy EEG signals is proposed.After the frequency bands were divided by wavelet packet transform,the mixed features of each rhythm were extracted.A total of 23 features,which were extracted in feature extraction step,were taken as classification data after normalization.(3)Dimensionality reduction.It would bring dimension disaster and lead to over fitting of classifier in the situation that the high-dimensional data was directly used as classifier to train data.Therefore,it’s necessary to reduce the data dimension.Linear and nonlinear dimensionality reduction methods were introduced to analyze the dimensionality reduction of the feature matrix.The simulation results showed that the isometric mapping(ISOMAP)algorithm had the best dimensionality reduction effect.(4)Classifier design based on ensemble learning algorithm.After dimensionality reduction,the feature information was used as the sample to train the classifier.The idea of integration was introduced to improve the generalization ability of the classifier.Support vector machine,Naive Bayes and K-nearest neighbor algorithm were chosen as base classifiers to construct a heterogeneous classifier,which was combined with back-propagation neural network in decision method.In the experiment,the accuracy of the heterogeneous classifier combined with backpropagation neural network reached 99.44% in binary classification,97.04% in three-class classification,and 88.67% in five-class classification.The accuracy was much higher than that of the base classifier and the heterogeneous classifier with voting method.